Research On Integration Of SAR Target Detection And Recognition Based On Deep Learning

There is a significant difference between interpretation of synthetic aperture radar(SAR)images and that of optical images because of unique imaging method and influence of special noise.Traditional SAR automatic target recognition(SAR-ATR)systems have many limitations,such as procedure isolation and error amplification,which make it hard to realize a truly automatic and integrated SAR interpretation system.In this paper,after introducing components and limitations of traditional SAR-ATR systems,we analyze and apply deep learning methods into field of SAR image processing,then realize an integrated model that can detect and recognize targets in SAR images at once,especially for targets in large-scale SAR images.First of all,unique features of SAR images are acquired;research on methods and structures of traditional SAR-ATR system,such as MIT SAR-ATR and MSTAR SARATR,are carried out.Then we explore application of some normal machine learning methods in SAR detection and recognition.Performances of these models are evaluated and exhibited,which illustrate the effectiveness of machine learning methods in field of SAR image interpretation.These models are also employed in our comparison experiment.Then we employ convolutional neural network(CNN)to carry out research on deep learning methods.We have explored structure,key functions,training tricks as well as ways of realization of this integrated SAR detection and recognition model based on CNN.Then interpretation accuracy,anti-noise and translational invariant performance of this model in SAR target recognition based on comparison experiments are evaluated.At last,after researching and analyzing range of use,advantages and disadvantages of some widely cited segmentation methods,we put forward a self-adapting segmentation algorithm and non-maximum suppression among regions(NMSR)to realize integrated interpretation model of large-scale SAR images.We also exhibit performance of this integrated SAR interpretation model on large-scale images with different background,and analyze results in the comparison experiments.The results show that the model is effective and efficiency.